Diver Segmentation Frames for Diving48

This dataset contains 303 segmentation labeled frames with diver segmentation from 303 randomly chosen videos (1 frame per video) from the training split of the Diving48 dataset. When there are 2 divers in the video, they are segmented as separate instances. 55 frames contains 2 diver instances, meaning that there are 358 instances in total. This dataset also contains a trained instance of MaskRCNN which has been fine-tuned to this dataset, and which is used in the associated publication. The trained model achieves .931 box IoU on a random validation set of 13 frames from this same collection of labeled frames (search for 'checkpoint' below to find the model checkpoint among the image files). Note that the Statistical Visual Computing Lab in San Diego (http://www.svcl.ucsd.edu) has the copyright to the Diving48 dataset. Please cite the RESOUND paper, if you are using any data related to the Diving48 dataset, including our labeled frames here "RESOUND: Towards Action Recognition without Representation Bias", Li et al., ECCV 2020. Please also cite our paper "Recur, Attend or Convolve? Frame Dependency Modeling Matters for Cross-Domain Robustness in Action Recognition", Broomé et al., arXiv 2021, if these frames are useful for you.